Gradient-Free SNN Training via Low-Rank Evolution Strategies
A new method for training Spiking Neural Networks (SNNs) without gradients has been introduced, using a low-rank factorization of Evolution Strategies (ES) perturbations called EGGROLL. SNNs are energy-efficient on neuromorphic hardware but are hard to train due to the non-differentiable spike threshold. Surrogate-gradient methods require backpropagation, which is incompatible with on-chip learning. ES offers a gradient-free alternative but scales poorly with parameter count. EGGROLL reduces per-generation memory from O(mn) to O(r(m+n)). Tested on a Leaky Integrate-and-Fire SNN with N-MNIST dataset, it achieved 79.21% test accuracy and reduced wall-clock time by 2.23x compared to full-rank ES. The paper is available on arXiv.
Key facts
- Method called EGGROLL uses low-rank factorization of ES perturbations
- Reduces memory from O(mn) to O(r(m+n))
- Achieved 79.21% test accuracy on N-MNIST
- Wall-clock time reduced by 2.23x relative to full-rank ES
- Trained Leaky Integrate-and-Fire SNN
- Gradient-free approach compatible with on-chip learning
- Addresses non-differentiable spike threshold in SNNs
- Published on arXiv with ID 2605.30361
Entities
Institutions
- arXiv